By Yvanne Enever, Founder & CEO of PHARMExcel
The wearable health device market has moved a long way past step counters. Devices that sense, interpret, and sometimes act on physiological data using artificial intelligence are now part of active clinical research, from consumer fitness trackers estimating blood pressure to experimental brain-monitoring platforms in NHS-sponsored trials.
This development in tech has forced regulators on both sides of the Atlantic to rethink how they classify, evaluate, and monitor these tools, since a device that continues to learn after it reaches the market does not fit neatly into the current frameworks built for static hardware.
A regulatory landscape in motion
In the UK, the Medicines and Healthcare products Regulatory Agency (MHRA) has been running the AI Airlock, a regulatory sandbox launched in spring 2024 specifically for artificial intelligence as a medical device. Rather than applying existing rules to novel AI systems from a distance, the Airlock brings manufacturers, the MHRA, NHS teams, and the UK’s Approved Body consortium together to test real AI-enabled devices in controlled, real-world-like conditions before they reach full market authorisation.
Early cohorts have included generative AI tools used in radiology reporting, and a second phase running into 2026 has extended the work to large language models, voice-based tools, and specialised diagnostics for cancer and rare disease.
The Department of Health and Social Care recently committed £1.2 million a year through to 2029 to move the programme beyond short-term pilots, with findings feeding into a National Commission tasked with recommending a fuller AI-in-healthcare regulatory framework.
Meanwhile, the US has been moving in a different direction. In January 2026, the FDA issued revised final guidance on general wellness products and clinical decision support software, both of which touch AI-enabled wearables.
The updated guidance widens the range of non-invasive wearables, including ones estimating blood pressure or blood glucose, that can be marketed as general wellness products rather than regulated medical devices, provided they stick to wellness claims and avoid diagnostic or treatment claims.
Software that assists but does not autonomously drive clinical decisions is similarly given more room to operate outside formal device clearance. The agency has framed this as removing unnecessary friction for genuinely low-risk products, while stressing that the underlying safety standards for higher-risk, disease-related claims haven’t changed.
Taken together, the UK and US approaches illustrate two different strategies for the same underlying problem: the UK is building dedicated testing infrastructure for higher-risk AI systems before they reach the market, while the US is narrowing the definition of what counts as a regulated device in the first place.
Neither approach is settled; both are explicitly described by their respective regulators as “work in progress”, with further guidance and legislative recommendations expected through 2026 and beyond.
How this affects live clinical trials
Regulatory sandboxes and guidance documents matter most when they meet an actual study design, and it’s a question we deal with directly in our own work. At PHARMExcel, we are currently managing one such example: a UK neuroscience research programme run in partnership with Barking and Havering NHS Trust, delivering a feasibility and early-safety study of a novel brain-machine interface.
The platform, developed by the US-based Forest Neurotech and known as the Ultrasonic Neural Interface, uses focused ultrasound rather than the implanted electrodes typical of older brain-computer interface approaches, aiming to image and interrogate activity across the whole brain rather than a single region.
The study, funded by the UK’s Advanced Research and Invention Agency (ARIA), recruits participants who have previously undergone a decompressive craniectomy, since the resulting gap in the skull allows ultrasound to reach brain tissue that intact bone would otherwise block.
As an investigator-led, academic-sponsored programme working with a genuinely novel research platform, this is exactly where regulatory frameworks are least settled: it is not a consumer wearable seeking a wellness exemption, but a Class IIb investigational device whose long-term regulatory pathway is still being defined alongside the research itself.
As the CRO managing this trial, our work includes ethics and Health Research Authority coordination, participant safety monitoring, and quality oversight designed to satisfy NHS R&D governance requirements from the outset, rather than retrofitting compliance once a device’s classification is decided.
It’s a useful illustration of why feasibility studies for novel technology tend to move more carefully than device validation trials in more established categories, and why sponsors increasingly look for research partners with specific experience in this space rather than generalist trial management.
The safety and ethics questions that don’t go away
Loosening oversight for low-risk wearables and building sandboxes for higher-risk AI systems both leave a set of harder questions unresolved. The MHRA’s own reporting from its AI Airlock pilots has flagged several recurring issues: how to manage AI systems that continue to adapt after deployment rather than staying fixed at the point of approval, how to ensure an AI tool’s outputs are properly grounded in verified clinical information rather than plausible-sounding error, and how to detect clinicians or patients over-relying on an AI recommendation rather than exercising independent judgement.
None of these are solved by a single approval decision; they require ongoing post-market monitoring, which is a different regulatory muscle than the one built for traditional, unchanging devices.
There is a parallel tension on the wearables side of the FDA’s new guidance. A device marketed purely for “wellness” is, by definition, not meant to be relied on for diagnosis or treatment decisions. But when a wrist-worn sensor reports a blood pressure or blood glucose reading using the same units and thresholds a clinician would use, the practical line between “wellness information” and “something I should act on medically” is set by the user, not the label.
Regulators on both sides of the Atlantic have acknowledged this indirectly, with the FDA’s updated guidance addressing how alerts and notifications should be designed so they don’t imply a clinical judgement the device isn’t authorised to make.
For patients and the public, the direction of travel is something of a double-edged sword: broader access to AI-enabled monitoring and faster pathways for genuinely useful tools, set against a regulatory system that is still building the infrastructure to catch problems that only appear after a device is in continuous, real-world use.
That is arguably the more interesting regulatory story here; not whether AI-enabled wearables will be approved, but what happens to safety oversight in the months and years after they are.
Disclosure: Yvanne Enever is Founder and CEO of PHARMExcel, a contract research organisation involved in the clinical research programme discussed in this article.
Disclaimer: This article is provided for general informational and educational purposes only and does not constitute medical, clinical, legal or regulatory advice. Regulatory requirements for artificial intelligence, wearable technologies and medical devices are evolving and may vary by jurisdiction, device classification and intended use. Information was considered current at the time of publication but may change as regulators issue new legislation, guidance or interpretations. References to organisations, companies, products, technologies or clinical studies do not constitute endorsement by Open MedScience. Readers involved in the development, investigation, approval or use of medical devices should consult the relevant regulatory authorities and suitably qualified professional advisers for guidance applicable to their circumstances.
home » blog » medical technologies »
